Kirk L. Duffin

Distributed and hardware accelerated computing for clinical medical imaging using proton computed tomography (pCT)

Proton computed tomography (pCT) is an imaging modality that has been in development to support targeted dose delivery in proton therapy. It aims to accurately map the distribution of relative stopping power. Because protons traverse material media in non-linear paths, pCT requires individual proton processing. Image reconstruction then becomes a time-consuming process. Clinical-use scenarios that require images from billions of protons in less than ten or fifteen minutes have motivated us to use distributed and hardware-accelerated computing methods to achieve fast image reconstruction. Combined use of MPI and GPUs demonstrates that clinically viable image reconstruction is possible. On a 60-node CPU/GPU computer cluster, we achieved efficient strong and weak scaling when reconstructing images from two billion histories in under seven minutes. This represents a significant improvement over the previous state-of-the-art in pCT, which took almost seventy minutes to reconstruct an image from 131 million histories on a single-CPU, single-GPU computer.

X-ray speckle visibility spectroscopy in the single-photon limit.

The technique of speckle visibility spectroscopy has been employed for the measurement of dynamics using coherent X-ray scattering. It is shown that the X-ray contrast within a single exposure can be related to the relaxation time of the intermediate scattering function, and this methodology is applied to the diffusion of 72 nm-radius latex spheres in glycerol. Data were collected with exposure times as short as 2 ms by employing a resonant shutter. The weak scattering present for short exposures necessitated an analysis formalism based on the spatial correlation function of individual photon charge droplets on an area detector, rather than the usual methods employed for intensity correlations. It is demonstrated that this method gives good agreement between theory and experiment and thus holds promise for extending area-detector-based coherent scattering methods to the study of faster dynamics than previously obtainable.

Fast focal length solution in partial panoramic image stitching

Accurate estimation of effective camera focal length is crucial to the success of panoramic image stitching. Fast techniques for estimating the focal length exist, but are dependent upon a close initial approximation or the existence of a full circle panoramic image sequence. Numerical solutions of the focal length demonstrate strong coupling between the focal length and the angles used to position each component image about the common spherical center. This paper demonstrates that parameterizing panoramic image positions using spherical arc length instead of angles effectively decouples the focal length from the image position. This new parameterization does not require an initial focal length estimate for quick convergence, nor does it require a full circle panorama in order to refine the focal length. Experiments with synthetic and real image sets demonstrate the robustness of the method and a speedup of 5 to 20 times over angle based positioning.

An Analysis of a Distributed GPU Implementation of Proton Computed Tomographic (pCT) Reconstruction

Proton computed tomography (pCT) is an imaging modality being developed to support targeted dose delivery in proton therapy. It aims to accurately map the distribution of relative stopping power in the imaged body. Because protons traverse material in non-linear paths, pCT requires individual proton processing and image reconstruction becomes a time-consuming process. We discuss the transformation of image reconstruction techniques from single CPU/GPU implementations to create a hybrid multi-CPU/GPU approach. We demonstrate a reduction of computation time from almost 7 hours down to 53 seconds.

Performance Model Selection for Learning-based Biological Image Analysis on a Cluster

Microscopic images with increased scale and content call for high performance computing when applying automatic tools for biological image analysis. Speed of analysis can be improved at various stages. In learning-based models, selecting suitable algorithms for a given problem can be a lengthy process given the large pool of algorithms and the variety of biological problems. In this paper, we describe a portable method for efficiently and adaptively selecting an effective model for biological image classification as a step toward the goal of achieving high throughput biological image analysis. We implemented a high performance tool which extends the bioimage classification and annotation platform BIOCAT by deploying the model selection process on a cluster using a distributed design based on remote method invocation. The high performance model selection, when tested and compared using ten benchmarking data sets, is shown to not only dramatically increase the speed of the learning process, but also bring improved accuracy to several state-of-the-art data sets for bioimage classification. These achievements are attributed to the combined power of BIOCATs adaptive model selection as well as the capability of distributed model evaluation. The tool is deployable to various types of distributed environments.